This content has been automatically translated from Ukrainian.
Artificial intelligence can confidently explain complex physical theories, help write code, or summarize the content of a book. And in a minute — name an article that never existed, attribute a fabricated quote to a famous person, or "remember" a non-existent event. Moreover, it can do this so convincingly that the mistake is not always easy to notice.
Such cases are called AI hallucinations. The term is somewhat conditional: artificial intelligence does not experience hallucinations in the human sense. It refers to situations where a generative model creates false, fabricated, or unfounded information and can present it as if it were credible.
In documents from the American National Institute of Standards and Technology (NIST), for example, the term confabulation is used for this phenomenon, while "hallucination" is mentioned as a common name for such errors.
What "AI hallucination" means
An AI hallucination, or AI hallucination, is a model's response that may appear quite plausible but contains statements that do not correspond to the facts, data, or context of the query.
For example, AI can:
- invent the title of a scientific study;
- name a non-existent author of a book;
- add an event to a person's biography that never happened;
- create a false or non-existent reference;
- misrepresent the content of a document;
- attribute a certain saying to the wrong person;
- confidently state an incorrect date or number.
The peculiarity of such errors is that the model does not necessarily demonstrate uncertainty. A fabricated fact can be formulated as smoothly and convincingly as a correct one.
That is why hallucinations remain one of the significant problems of generative AI even in modern language models.
Why AI invents things at all
To understand the nature of hallucinations, one needs to grasp how large language models work.
During basic training, such a model learns to predict the next fragments of text — tokens — based on the previous context. It does not simply choose one "most likely word," but evaluates the probabilities of possible continuations. Modern systems are additionally trained to follow instructions, work with different types of queries, and in some cases, connect to searches, databases, and other tools.
However, the very principle of text generation does not guarantee the truth of every generated statement. As NIST notes, statistical predictions can yield both factually correct and consistent results, as well as incorrect or internally contradictory ones.
For example, the model learns well how a bibliographic reference usually looks: authors' names, title of the work, journal, year, DOI number. Therefore, it can generate a record that outwardly is almost indistinguishable from a real one. But a specific article with that title may never exist.
In other words, the model can reproduce the form of a correct answer very well, without having sufficient grounds for all the facts within it.
Hallucination is not just any mistake
The boundary between concepts is not always clear, but "hallucination" usually refers not to every incorrect AI response, but to cases where the system generates false information without sufficient grounding in facts or input data.
If the model makes an arithmetic mistake, that is primarily a mathematical error.
However, if it states:
“In 2019, Professor Anna Miller published a study in Nature about…”
and neither such a researcher nor a publication actually exists, this is a typical example of a hallucination.
Another common case arises when working with documents. The text may not contain an answer to the user's question, but instead of the phrase "this is not mentioned in the document," the model creates a logical and plausible explanation on its own.
Why hallucinations often seem true
It is usually easier for a person to suspect a mistake if the interlocutor hesitates, gets confused, or contradicts themselves. This does not always work with generative AI.
An incorrect statement can appear in a well-constructed sentence, be accompanied by explanations, and fit well into the rest of the text. The form of the response creates a sense of competence, even though the information itself may be false.
Especially convincing are fabricated quotes, references, and scientific sources. The model can create a statement that aligns with the views of a well-known person or a study title that perfectly fits the topic. Sometimes real and fabricated details are mixed within one text: the journal exists, the topic sounds logical, but the authors or specific publication do not.
Such cases are particularly difficult to notice. In a large response, almost everything can be correct, while only one date, quote, or title may be fabricated.
When the risk of hallucinations is higher
The risk depends on the specific model, query, and how it operates, but one should be especially cautious when very precise or obscure facts are required.
These can be verbatim quotes, pages of books, bibliographic data, little-known historical events, document numbers, exact statistics, or information about a very narrow topic. A separate problem is current information: if the system does not have access to fresh sources, it may not be aware of recent changes.
A hallucination can also be triggered by the question itself if it contains an incorrect assumption.
For example:
“Why did Albert Einstein receive the Nobel Prize for the theory of relativity?”
The formulation of the question is incorrect. The Nobel Prize in Physics for 1921 was awarded to Einstein "for his services to theoretical physics, and especially for the discovery of the photoelectric effect." The theory of relativity is not mentioned in the official motivation for the prize.
If the model simply accepts the user's assumption as a fact, it may begin to explain in detail something that did not actually happen.
Can hallucinations be completely eliminated
AI developers are trying to reduce the number of such errors: models are trained to better respond to uncertainty, work with external sources, and not provide answers where there is insufficient information.
One common approach is RAG, or Retrieval-Augmented Generation, meaning generation augmented by information retrieval. In such a system, the language model can rely not only on knowledge acquired during training but also on documents found in external storage when forming a response. The RAG approach was described by researchers in 2020 as a combination of the parametric memory of the language model with access to external non-parametric memory.
However, even access to sources does not make the system infallible. The model may misunderstand the material found, use an irrelevant source, confuse data, or draw an unfounded conclusion.
Therefore, hallucinations can be reduced, but currently, the problem should not be considered fully resolved.
How to recognize a possible hallucination
The most important rule is not to perceive the confident tone of AI as proof of correctness.
It is especially worth checking specific dates, statistics, quotes, research titles, references, and other information where even a small error can significantly change the content of the response.
A suspicious signal can also be an extremely precise answer to a very narrow question. If the model names a page of a little-known book, a verbatim quote, or a document number, it is better to open the original source and verify that this data actually exists.
You can ask the AI to name the source, but this alone does not guarantee reliability: the model can fabricate not only the fact but also the source that supposedly confirms it.
It is more reliable to check the document, scientific publication, or official page directly or use a system that shows the sources on which the response is based.
Do hallucinations mean AI cannot be trusted
Hallucinations do not render generative AI useless. They rather show why it is important to understand the limits of such systems.
If you need to come up with title options, structure text, explain concepts in simple terms, or generate ideas, AI can significantly speed up the work. A different situation arises when its response is used as a source of precise facts.
AI is better perceived not as an infallible encyclopedia but as a tool for working with information. It can quickly explain, compare, summarize, and help find connections, but important statements should still be verified.
That is why the phrase “AI can make mistakes” is not a formal warning but an important disclaimer. Generative models have learned to create plausible responses extremely well, but plausibility and reliability are not always the same.
